DFKI AI Next 3 englisch

AI Innovation Twinned AI, Trusted AI, and - Authors: Prof. Dr. Antonio Krüger, Prof. Dr. Philipp Slusallek, Dr. Frederic Stahl digital twin For the ﬁrst time since the breakthrough in generative AI, artiﬁcial intelligence has been able to gener- ate creative content. This development is based on very large data collections and deep neural networks, particularly large language models (LLM). In principle, robots have the ability to ac- cess the entire knowledge of an LLM and use it to gener- ate actions. In the digital twin – a virtual representation of the physical world constructed on available data – the robot‘s virtual alter ego already has the ability to act and move autonomously. What is still absent in AI‘s ability to transfer this to action in the real world is the enrichment of the digital twin with the dynamic data from the physical environ- ment and a concept with which the robot can harmonize its knowledge of the world from the LLM with the aware- ness of its current environment. The Twinned AI innovation ﬁeld at DFKI now aims to achieve this by placing robots in the physical world and enabling them to act in a context-adaptive manner. Ro- bots oriented in this way could be used wherever they interact with humans, i.e., in unstructured environments with changing parameters such as everyday assistance or caregiving situations. In an industrial setting, Twinned AI offers the op- portunity to open up manufacturing to a large extent for human-robot collaboration. The end vision is of a future factory designed as a neutral platform on which any product can be manufactured with robotic support. for innovation The expanding integration of AI into major sectors of the economy and society requires robust mechanisms to en- sure acceptance and trust. Adjectives such as reliable, transparent, fair, robust, inclusive, and ethically respon- sible are the measures used to describe the basis for trustworthy AI systems. Researchers at DFKI are working with industry, science, and government in applying these measures to deﬁne clear test criteria for AI systems. 4

AI Innovation These criteria then serve as the means to guarantee the development of trust- worthy systems. The aim is to create the foundation for solid trust in the use of AI, i.e., technological innovation with social re- sponsibility. Trustworthy AI is a priority of DFKI’s MISSION KI project and also the focus of the CERTAIN initiative (Center for European Research in Trusted AI). MISSION KI supports the development of uniform testing and certiﬁcation methods for AI applications and systems. In collaboration with partners such as TÜV (English: Technical Inspection Association), the project helps deﬁne transparency, reliability, and ethical re- sponsibility. The output serves as a basis for regulatory requirements and offers economic security for domestic AI companies. CERTAIN develops and promotes new technology standards and AI governance that guarantee AI systems will act reliably, safely, and ethically. As a collective Euro- pean initiative, CERTAIN brings together various efforts and stakeholders from research, industry, administrative, and public organizations. DFKI Darmstadt focuses on AI safety for large AI mod- els such as LLMs. Research focuses on how unintentional damage caused by AI systems can be prevented using targeted methods and frameworks. This creates the basis for robust and safe AI applications. Twinned AI and Trusted AI require powerful and con- stantly growing foundation models that, inevitably, go hand in hand with increasing energy needs. Sustainable AI is dedicated to the question of how AI systems can become more resource-efﬁcient and contribute to sus- tainable economic development. The industry lacks standardized tools to make the technology‘s impact transparent and measurable, mak- ing it difﬁcult to reconcile AI innovation with ecological and social goals. DFKI takes on this challenge with the DFKI4Planet Competence Center, which is committed to sustainable AI solutions. This is not only about sustainable appli- cations but also about the sustainability of AI itself. AI must be measurable, fair, and resource-efﬁcient. DFKI is developing the tools necessary to evaluate the energy consumption, emissions, and resource consumption of AI technologies. These tools provide decision-making aids that help companies beneﬁt from the sustainable and efﬁcient use of AI. Sustainability becomes a driver of innovation while ensuring long-term com- petitiveness and future vi- ability. Companies that use sustainable AI solutions reduce costs, meet legal requirements, and position themselves as pioneers for responsible technologies. bundled The ap- proach of DFKI makes a decisive contri- bution to the future development of not only powerful and understandable AI technologies but also ensures its safe and responsible use. The further development of artiﬁcial intelligence relies on taking a comprehensive approach based on the close rela- tionship among these three innovation streams: Twinned AI, Trusted AI, and Sustainable AI. The combination of physical integration, trustworthiness, and sustainability not only creates technological added value but also ena- bles the responsible design of future AI systems to meet the needs of society. 5

Interview seamless human-robot - - Professor Krüger, what is the main idea behind Twinned AI? The aim of Twinned AI is to develop new functional possibilities for AI technologies and seamlessly inte- grate intelligent robotic systems and the physical world. New approaches to the development and implemen- tation of such systems are needed to achieve this. AI systems must be able to respond flexibly when en- countering unknown terrain. Ideally, they will be able to recognize their environment, assess the capabilities of other actors (human or machine), and adapt their own actions accord- ingly, all in real-time. At its core, the idea is to make exist- ing concepts and knowledge tangi- ble through testing and veriﬁcation using real-time data from the envi- ronment. This implies the dynamic creation of a “wealth of experience” that is then used to influence future decisions and actions. Any change in the real world has a direct impact on the virtual model. By continuous- ly combining real-world data with simulations, AI seamlessly inte- grates the two to create an adaptive learning system. What impact does a seamless connection between virtu- al and real worlds have on industrial production? Industry 4.0 has made artiﬁcial intelligence and the digital twin integral parts of the digital trans- formation. The digital twin maps production processes into a virtual environment and serves as a pow- erful planning tool for the efﬁcient deployment of machines, materi- als, and workforce. Resources are conserved, and other advantages beneﬁt many companies where ad- vanced plant designs and individual process steps can be simulated and optimized. Supply chain problems are mitigated, and alternative pro- cess designs compensate for staff- ing shortages. On the other hand, the data only flows in one direction. No matter how large the database of machine parameters, runtime information, and sensor data, the digital twin is ultimately simply a reflection of reality. Twinned AI is the next step, a bridge over the gap between the virtual and real worlds. 6

The robotic system continuously synchronizes both levels, mapping the physical environment in the digital twin in real-time. What is necessary to enable AI in production systems to move beyond simulation and process optimization? Until now, the digital twin has not had any interface to real-time movements and processes that would enable autonomous and environmentally-aware robotic actions. Ultimately, the aim of Twinned AI is to support bidirec- tional interaction between the real environment and the digital mo- del – taking into account real- time environmental data, physical forces, and natural laws. Twinned AI creates the opportunity for mixed teams and human-robot collaboration in an unstructured production environment. What research challenges still have to be overcome? If machines are to become truly capable of autonomous action, a contextual basis is required for more generic AI models. One major challenge is to combine the central concepts of digital twins and dual reality with powerful foundation models. The aim is not only to generate textual or visual content but also to create speciﬁc com- mands for direct implementation in the physical world. Furthermore, Interview large language and action models must be combined with knowl- edge-based planning methods. The path to developing robust, stable, and neuro-explicit AI systems involves connecting data-driven generative AI to explicit mechanis- tic models. This provides a basis for the development of robotic systems that act safely, flexibly, and appro- priately in the real world. What will future industrial at companies that make use of Twinned AI. One potential vision of the future depicts the factory as a neutral, highly flexible platform on which a wide variety of products can be manufactured aided by ro- bots. Twinned AI takes human-ro- bot collaboration to a new level: industrial production features an increasingly adaptive manufactur- ing environment in which humans, machines, and AI are seamlessly working together. production look like when Twinned AI is implemented? New, flexible, and intelligent auto- mation systems become possible 7

Twinned AI AI-supported Narrow, precise gaps on doors, hoods and tailgates are considered a quality feature of modern vehicles. Precise gap dimensions are not de- signed for simply esthetic reasons, but also to ensure important func- tions – for example, a uniform clos- ing force and optimum tightness. A great deal of effort is expended in the production process on getting the correct door alignment. (AI- DFKI Project AIQUAMA Based Quality Management for Smart Factories) has set up an ex- perimental production environment to investigate the quality manage The worker receives AI-based instruc- tions for an exact adjustment of the door via a projection. ment of production processes for car body construction. The project team develops software to create a detailed view of the production environment using intrinsic sensor data from the production plant along with camera imagery. The data also includes the people involved and their current status. The aim is to minimize the effort involved in get- ting the proper door alignment and, at the same time, ensuring quality. The system recognizes, for ex- ample, whether a car door is open or closed, whether a worker is working on a door, or is on the way to do it. All the raw data collected is pro- cessed in the administration shell, where the actual status is compared to target status. If the gap dimen- sion needs correcting, the worker receives precise instructions from the Industrial Metaverse – either via mixed reality glasses or as a screen projected directly onto the vehicle door. The worker then proceeds to readjust the door screw connection in a targeted manner. The DFKI experimental production scenario is closely linked to the in- novation stream of Twinned AI. Twinned AI enables bidirection- al communication between digital models and the physical environ- ment – incorporating real-time en- vironmental data, physical forces at play, and the laws of nature. Twinned AI extends the concept of digital twins in Industry 4.0 to ad- vanced intelligent models that not only simulate and optimize physical processes, but can actively control and execute industrial production steps. The AIQUAMA production scenario is the ﬁrst stage in this de- velopment – from pure data collec- tion and analysis to automated ver- iﬁcation and recommendations for speciﬁc action. AIQUAMA is funded over a period of three years by the Federal Minis- try of Education and Research with around 1.8 million euros. The aim is to detect anomalies and defects in the production process at an early stage to save the high-costs rework- ing, time-consuming ﬁnal inspec- tions, and delivery delays. In addi- tion to DFKI, project partners include the Czech Technical University in Prague, the CEIT at the University of Brno, the Technical University of Os- trava, as well as application partners Volkswagen AG and Škoda Auto. 8

Trusted AI innovation Artiﬁcial intelligence is increasingly transforming soci- eties and industries by opening up immense opportuni- ties as well as exposing key challenges. To fully realize the potential of AI, technological innovation must align with social responsibility. Speciﬁcally, this means the necessary public trust must be established and main- tained. Working together with partners from industry and science, DFKI researchers are building the founda- tions to make trusted AI accessible throughout Europe. Transparency is a key building block in this effort. Users must be able to follow how an AI system arrives at decisions – especially, those involving a safety- or ethically-related context. To this end, DFKI is working with industry partners to develop the test criteria and certiﬁcation procedures that deﬁne minimum stand- ards that permit methodical testing of system traceabil- ity and reliability. Applicable to cross-sector industries, these standards and procedures provide the necessary orientation and support for reliable implementation while enabling guarantees for AI systems. The current research focus, in addition to certiﬁca- tion, is on system validation. DFKI is developing test procedures for fairness, transparency, and robustness. Models are subjected to simulations and deductive testing in realistic scenarios to identify weaknesses at an early stage. This not only makes AI models more ef- ﬁcient but also more understandable and reliable. Technical excellence alone is not sufﬁcient to gain acceptance and establish a basis of trust. Binary per- formance measurements of AI systems must be sup- plemented with practical beneﬁts for people. The basis that allows an AI to act as a cognitive sparring partner and make complex decision-making understandable requires narrative, explanatory models, and transpar- ent interactions. International AI safety report released with DFKI CEO Antonio Krüger was appointed by the Fede- ral German Government to participate as the German representative on an international panel of experts tas- ked to prepare and publish the International Scientiﬁc Report on the Safety of Advanced AI. The panel’s report presents the opportunities and chal- lenges of increasingly powerful AI systems and discus- ses possible solutions. In particular, the report analyzes the development of so-called “general-purpose AI” – versatile AI systems undergoing rapid advances in capa- bilities. The experts do not lose sight of the possibilities offered by these technologies while also pointing out the perils, such as potential malfunctions. The report was a preparatory work for the AI Action Summit on February 10 and 11, 2025, in Paris. www.dfki.de/en/web/news/ai-safety-report Many DFKI-wide projects have already shown that re- sponsible innovation is not an oxymoron: to the con- trary, the terms are mutually dependent. A structured approach that reflects the needs of society, transparen- cy, and certiﬁcation can propel Europe into a pioneering role in the development of trustworthy AI, helping to shape a future in which AI is efﬁcient, understandable, and socially responsible. 9

Sustainable AI in the baked goods Millions of tons of food end up as waste in Germany every year – a signiﬁcant percentage of it orig- inates in- dustry. Brammibal‘s Donuts, a Berlin-based vegan pastry manu- facturer, is doing something about it by implementing an AI solution. Working together with Foodtracks, a software vendor to bakeries and organic supermarkets from Mün- ster, the company wants to con- serve resources and reduce food waste. The pilot project launched by Green-AI Hub Mittelstand is de- veloping the solution together with DFKI’s DRX research department (Design Research eXplorations). AI algorithms predict the daily demand for baked goods and reduce the overpro- duction of perishable food. fluctuations and regional differenc- es. DFKI Berlin helped to optimize the models in combination with intelligently designed interfaces. Empowered to make more precise ordering decisions, store managers quickly gained conﬁdence in the AI-supported forecasts. The solution implemented at Brammibal’s Donuts has the po- tential to save more than 200,000 donuts a year from being overpro- duced. Foodtracks reduced bakery waste in all of its customers’ 2,800 stores by up to 3,500 tons within a year. The Green-AI Hub Mittelstand is pioneering the use of AI for great- er resource efﬁciency and savings. It is speciﬁcally aimed at SMEs – practical, solution-oriented, and directly on site. The BMUV initiative is coordinated by Zukunft Umwelt Gesellschaft (ZUG) and implement- ed jointly by DFKI, the Wuppertal Institute, and the VDI Technology Center. The Green-AI Hub Mittelstand is an AI initiative of the Federal Ministry for the Environment, Na- ture Conservation, Nuclear Safety and Consumer Protection (BMUV), which implements practical pro- jects for small and medium-sized enterprises (German: Mittelstand) in partnership with DFKI. AI algorithms developed in the pilot project with Brammibal’s and Foodtracks can predict dai- ly demand for baked goods more precisely. Fresh products such as Brammibal’s donuts are pro- duced without preservatives, which makes them perishable and, espe- cially, subject to production errors. Production planning is a challenge as demand fluctuates greatly de- pending on the day of the week, lo- cation, and customer preferences. The project aims to reduce both the amount of waste and the envi- ronmental impact of the manufac- ture of baked goods. Foodtracks already had models that were fur- ther developed and, in part, sup- plemented with new models to incorporate additional external data that better reflects seasonal 10

About us Intelligent solutions for the DFKI has more than 35 years of experience in human-centric AI, conducting research in key areas of basic research and future- oriented applications, while always maintaining a focus on social relevance and scientiﬁc excellence in the ﬁeld of artiﬁcial intelligence. The German Research Center for Artiﬁcial Intelligence (DFKI) was founded in 1988 as a non-proﬁt public-private partnership. It operates facilities in Kaiserslautern, Saarbrücken, Bremen, Lower Saxony (Osnabrück and Oldenburg), Darmstadt, with labora- tories in Berlin, Lübeck, and a branch ofﬁce in Trier. It supports 29 research areas, ten competence centers, and eight living labs. Based on application-oriented research, it develops product functions, prototypes, and patentable solutions in the infor- mation and communication technology (ICT) sector. Funding is provi- ded by public funding sources and industrial development contracts. Project results and milestones are periodically reviewed institutionally and by a Scientiﬁc Advisory Board, an international panel of experts. In addition to the German Federal Ministry of Education and Research and the federal states of Rhineland- Palatinate, Saarland, Bremen, and Lower Saxony, several leading German and international high-tech companies from a wide range of industries also have seats on the board. www.dfki.de 11

Lübeck Oldenburg Bremen Berlin Osnabrück DFKI Site DFKI Laboratory Trier Darmstadt Kaiserslautern Saarbrücken Imprint dfki ai next; Issue March | 2025; Published by: German Research Center for Artiﬁcial Intelligence GmbH (DFKI); Trippstadter Str. 122, D-67663 Kaiserslautern; Phone: +49 631 20575 0; E-Mail: news@dfki.de; Editors: Heike Leonhard (responsible), Sandra Antakli, Andrea Fink, Jeremy Gob, Armindo Ribeiro; Layout: Lando Lehmann; One Vision Design Credits Cover: AI-generated image, Adobe Firefly; Page 3: DFKI, Oliver Dietze; Page 4: freepik; AI-generated image, Adobe Firefly; Page 5 top: AI-generated image, AI Lab, stock.adobe.com; Page 5 center and bottom: AI-generated image, Adobe Firefly; Page 7: DFKI, Oliver Dietze; Page 8: DFKI, Jaron Hollax; Page 10: Brammibal‘s Donuts GmbH; Page 12: DFKI.